{"course":{"slug":"ai-and-you","title":"AI, Safety and What Goes Wrong","url":"https://addaly.com/learn/ai-and-you","level":"beginner","track":"start"},"lesson":{"slug":"the-training-data-dispute","title":"The training data question, unresolved","url":"https://addaly.com/learn/ai-and-you/the-training-data-dispute","position":60,"of":73,"minutes":10,"answer":"Courts are separating unlawful acquisition from transformative use and weighing market substitution heavily, legislatures are diverging from Japan's broad permission to India's enumerated fair dealing, and no jurisdiction has settled the general question.","keyIdea":"Courts are separating unlawful acquisition from transformative use and weighing market substitution heavily, legislatures are diverging from Japan's broad permission to India's enumerated fair dealing, and no jurisdiction has settled the general question."},"body":{"format":"markdown","text":"## Do not expect an answer here\n\nWhether training a model on copyrighted work without permission is lawful is being decided right now, in several countries, under different laws, with rulings pointing in different directions.\n\n## The claim on each side\n\n**Rights-holders argue** that building a commercial product required copying their work; that the copying was at industrial scale; that some material was obtained from pirate sources; that models can reproduce protected expression; and that the output competes in the market for the originals, sometimes directly.\n\n**Developers argue** that training extracts uncopyrightable statistical patterns rather than expression; that the use is transformative because the model does something different from the original work; that no copy is distributed to users; and that requiring permission for every work would make the technology impossible, which is an argument about consequences rather than doctrine.\n\nBoth positions contain something true, which is why this is hard.\n\n## What courts have actually said\n\nA partial, deliberately non-conclusive list.\n\nIn **Thomson Reuters v. Ross Intelligence** (2025) a US court rejected a fair use defence where headnotes from a legal research service had been used to build a competing legal research tool. The tool was not generative, and the market-substitution argument was direct.\n\nIn **Bartz v. Anthropic** (2025) a US judge held that training a language model on books was transformative and qualified as fair use, while separately holding that downloading and retaining a library of pirated books was not, a split that separates the *use* from the *acquisition*. The case was subsequently reported to have settled for a very large sum in respect of the pirated-library claims.\n\nIn **Kadrey v. Meta** (2025) another US judge granted summary judgment to the developer on the record presented, while stating explicitly that the ruling did not establish that training on copyrighted work is generally lawful, and that a better-evidenced market-harm case might succeed.\n\nIn the **UK**, Getty Images' case against Stability AI proceeded to trial in 2025 with the principal training-related claims narrowed or dropped during proceedings, and the remaining judgment turned largely on trade mark and secondary infringement points rather than settling the training question.\n\nIn **India**, ANI's suit against OpenAI in the Delhi High Court, filed in 2024, is being watched closely, in part because India's fair dealing provisions are enumerated rather than open-ended, so a transformative-use argument of the American kind has less room to operate.\n\nRead those together and the pattern is: courts are distinguishing acquisition from use, weighing market substitution heavily, and declining to announce general rules.\n\n> Figure: The training-data question, as the courts have left it. Timeline: 2019, EU Digital Single Market Directive: text and data mining permitted, with a machine-readable opt-out for rights-holders; 2024, ANI files against OpenAI in the Delhi High Court; 2025, Thomson Reuters v Ross: fair use rejected where headnotes built a competing legal research tool; 2025, Bartz v Anthropic: training on books held transformative; retaining a pirated library held not to be; 2025, Kadrey v Meta: judgment for the developer on the record, expressly not a general rule; 2025, Getty v Stability (UK): judgment turns largely on trade mark; the training question stays unsettled. The pattern so far: acquisition is separated from use, market substitution weighs heavily, and no court has announced a general rule. India's enumerated fair dealing leaves less room for an American-style transformative-use argument.\n\n## The legislative side\n\n**Japan** has the most permissive position: Article 30-4 of its Copyright Act broadly permits use of works for information analysis where enjoyment of the expression is not the purpose, subject to limits.\n\nThe **EU** permits text and data mining under the 2019 Directive, with a research exception and a commercial exception that rights-holders may opt out of in a machine-readable way. The AI Act then requires general-purpose model providers to have a policy respecting those opt-outs and to publish a sufficiently detailed summary of training content. In practice the opt-out mechanism is contested: there is no single standard for expressing it and no easy way for a rights-holder to verify compliance.\n\nThe **UK** consulted in 2024–25 on a similar exception with a rights reservation, met sustained opposition from the creative industries, and the transparency question was fought over in Parliament during the passage of data legislation in 2025. The outcome remains unsettled.\n\nThe **US** has no legislation on this; it is being decided case by case.\n\n## What is reasonably safe to say\n\nFour things, stated at the level of confidence they deserve.\n\n**Acquisition matters separately from use.** Where copies were obtained unlawfully, that has been treated as a distinct wrong even by judges sympathetic to training as fair use. This is the clearest signal from the case law so far.\n\n**Market substitution is the decisive factor.** Where the output competes directly with the input (a tool that replaces the licensed product it learned from) defences have fared worse.\n\n**Verbatim reproduction is a separate problem.** Memorised output is expression, not pattern, and every framework treats it differently from training.\n\n**Your jurisdiction may reach a different answer**, and if you are relying on this in a business, that is a question for a lawyer in your country rather than an inference from an American headline.\n\n## What to do in the meantime\n\nIf you build on generated material commercially: prefer providers offering indemnities and read what they cover; keep records of what was generated when and with what; avoid prompting for named living artists or protected characters; and treat any output that looks like something you recognise as a warning rather than a success."},"check":{"question":"A US ruling found that training a language model on books was transformative fair use, while separately finding that downloading a pirated library was not. What is the significance of that split?","options":[{"text":"It separates how copies were obtained from their use: fair use cannot cure piracy","correct":true},{"text":"It establishes that training is lawful everywhere so long as the model is transformative","correct":false,"tempting":"Turns one district ruling into an international rule, when other judges in the same country declined to state any general rule."},{"text":"Developers may train freely provided they delete the pirated copies afterwards","correct":false,"tempting":"Deleting after the fact does not undo the copying already done, which is the act the second holding addressed."},{"text":"Fair use turns mainly on whether the works were published, and books are published","correct":false,"tempting":"Publication is part of the second fair use factor, but it was not what divided the two holdings."}],"why":"Training was the use and the download was the acquisition. The court judged each separately."},"prev":{"slug":"what-copyright-actually-protects","title":"What copyright covers, and what it never did","url":"https://addaly.com/learn/ai-and-you/what-copyright-actually-protects"},"next":{"slug":"licences-you-actually-have","title":"Reading the licence before you build on it","url":"https://addaly.com/learn/ai-and-you/licences-you-actually-have"},"authorship":"Founder and series editor: Dr Mustafa Mun. Written by AI models to a written standard. Source and assessment review continues. Academic review invited.","review":{"status":"Draft, open for academic review"},"cite":"Addaly (2026) 'The training data question, unresolved', in AI, Safety and What Goes Wrong. Edited by M. Mun. Addaly. Available at: https://addaly.com/learn/ai-and-you/the-training-data-dispute (Accessed: 21 November 2026).","licence":{"name":"CC BY-NC-SA 4.0","url":"https://creativecommons.org/licenses/by-nc-sa/4.0/"},"reader":{"schema":1,"title":"The training data question, unresolved","courseTitle":"AI, Safety and What Goes Wrong","module":{"index":8,"slug":"law-and-ownership","title":"Ownership, law and the public record"},"answer":null,"keyIdea":"Courts are separating unlawful acquisition from transformative use and weighing market substitution heavily, legislatures are diverging from Japan’s broad permission to India’s enumerated fair dealing, and no jurisdiction has settled the general question.","hero":{"kind":"figure","figure":{"t":"figure","label":"Figure","figure":{"caption":"The pattern so far: acquisition is separated from use, market substitution weighs heavily, and no court has announced a general rule. India’s enumerated fair dealing leaves less room for an American-style transformative-use argument.","events":[{"when":"2019","what":"EU Digital Single Market Directive: text and data mining permitted, with a machine-readable opt-out for rights-holders"},{"when":"2024","what":"ANI files against OpenAI in the Delhi High Court"},{"when":"2025","what":"Thomson Reuters v Ross: fair use rejected where headnotes built a competing legal research tool"},{"when":"2025","what":"Bartz v Anthropic: training on books held transformative; retaining a pirated library held not to be"},{"when":"2025","what":"Kadrey v Meta: judgment for the developer on the record, expressly not a general rule"},{"when":"2025","what":"Getty v Stability (UK): judgment turns largely on trade mark; the training question stays unsettled"}],"title":"The training-data question, as the courts have left it","type":"timeline"},"alt":"Figure: The training-data question, as the courts have left it. Timeline: 2019, EU Digital Single Market Directive: text and data mining permitted, with a machine-readable opt-out for rights-holders; 2024, ANI files against OpenAI in the Delhi High Court; 2025, Thomson Reuters v Ross: fair use rejected where headnotes built a competing legal research tool; 2025, Bartz v Anthropic: training on books held transformative; retaining a pirated library held not to be; 2025, Kadrey v Meta: judgment for the developer on the record, expressly not a general rule; 2025, Getty v Stability (UK): judgment turns largely on trade mark; the training question stays unsettled. The pattern so far: acquisition is separated from use, market substitution weighs heavily, and no court has announced a general rule. India’s enumerated fair dealing leaves less room for an American-style transformative-use argument."}},"blocks":[{"t":"h2","id":"do-not-expect-an-answer-here","runs":[{"text":"Do not expect an answer here"}]},{"t":"p","runs":[{"text":"Whether training a model on copyrighted work without permission is lawful is being decided right now, in several countries, under different laws, with rulings pointing in different directions."}]},{"t":"h2","id":"the-claim-on-each-side","runs":[{"text":"The claim on each side"}]},{"t":"p","runs":[{"text":"Rights-holders argue","b":true},{"text":" that building a commercial product required copying their work; that the copying was at industrial scale; that some material was obtained from pirate sources; that models can reproduce protected expression; and that the output competes in the market for the originals, sometimes directly."}]},{"t":"p","runs":[{"text":"Developers argue","b":true},{"text":" that training extracts uncopyrightable statistical patterns rather than expression; that the use is transformative because the model does something different from the original work; that no copy is distributed to users; and that requiring permission for every work would make the technology impossible, which is an argument about consequences rather than doctrine."}]},{"t":"p","runs":[{"text":"Both positions contain something true, which is why this is hard."}]},{"t":"h2","id":"what-courts-have-actually-said","runs":[{"text":"What courts have actually said"}]},{"t":"p","runs":[{"text":"A partial, deliberately non-conclusive list."}]},{"t":"p","runs":[{"text":"In "},{"text":"Thomson Reuters v. Ross Intelligence","b":true},{"text":" (2025) a US court rejected a fair use defence where headnotes from a legal research service had been used to build a competing legal research tool. The tool was not generative, and the market-substitution argument was direct."}]},{"t":"p","runs":[{"text":"In "},{"text":"Bartz v. Anthropic","b":true},{"text":" (2025) a US judge held that training a language model on books was transformative and qualified as fair use, while separately holding that downloading and retaining a library of pirated books was not, a split that separates the "},{"text":"use","i":true},{"text":" from the "},{"text":"acquisition","i":true},{"text":". The case was subsequently reported to have settled for a very large sum in respect of the pirated-library claims."}]},{"t":"p","runs":[{"text":"In "},{"text":"Kadrey v. Meta","b":true},{"text":" (2025) another US judge granted summary judgment to the developer on the record presented, while stating explicitly that the ruling did not establish that training on copyrighted work is generally lawful, and that a better-evidenced market-harm case might succeed."}]},{"t":"p","runs":[{"text":"In the "},{"text":"UK","b":true},{"text":", Getty Images’ case against Stability AI proceeded to trial in 2025 with the principal training-related claims narrowed or dropped during proceedings, and the remaining judgment turned largely on trade mark and secondary infringement points rather than settling the training question."}]},{"t":"p","runs":[{"text":"In "},{"text":"India","b":true},{"text":", ANI’s suit against OpenAI in the Delhi High Court, filed in 2024, is being watched closely, in part because India’s fair dealing provisions are enumerated rather than open-ended, so a transformative-use argument of the American kind has less room to operate."}]},{"t":"p","runs":[{"text":"Read those together and the pattern is: courts are distinguishing acquisition from use, weighing market substitution heavily, and declining to announce general rules."}]},{"t":"h2","id":"the-legislative-side","runs":[{"text":"The legislative side"}]},{"t":"p","runs":[{"text":"Japan","b":true},{"text":" has the most permissive position: Article 30-4 of its Copyright Act broadly permits use of works for information analysis where enjoyment of the expression is not the purpose, subject to limits."}]},{"t":"p","runs":[{"text":"The "},{"text":"EU","b":true},{"text":" permits text and data mining under the 2019 Directive, with a research exception and a commercial exception that rights-holders may opt out of in a machine-readable way. The AI Act then requires general-purpose model providers to have a policy respecting those opt-outs and to publish a sufficiently detailed summary of training content. In practice the opt-out mechanism is contested: there is no single standard for expressing it and no easy way for a rights-holder to verify compliance."}]},{"t":"p","runs":[{"text":"The "},{"text":"UK","b":true},{"text":" consulted in 2024–25 on a similar exception with a rights reservation, met sustained opposition from the creative industries, and the transparency question was fought over in Parliament during the passage of data legislation in 2025. The outcome remains unsettled."}]},{"t":"p","runs":[{"text":"The "},{"text":"US","b":true},{"text":" has no legislation on this; it is being decided case by case."}]},{"t":"h2","id":"what-is-reasonably-safe-to-say","runs":[{"text":"What is reasonably safe to say"}]},{"t":"p","runs":[{"text":"Four things, stated at the level of confidence they deserve."}]},{"t":"p","runs":[{"text":"Acquisition matters separately from use.","b":true},{"text":" Where copies were obtained unlawfully, that has been treated as a distinct wrong even by judges sympathetic to training as fair use. This is the clearest signal from the case law so far."}]},{"t":"p","runs":[{"text":"Market substitution is the decisive factor.","b":true},{"text":" Where the output competes directly with the input (a tool that replaces the licensed product it learned from) defences have fared worse."}]},{"t":"p","runs":[{"text":"Verbatim reproduction is a separate problem.","b":true},{"text":" Memorised output is expression, not pattern, and every framework treats it differently from training."}]},{"t":"p","runs":[{"text":"Your jurisdiction may reach a different answer","b":true},{"text":", and if you are relying on this in a business, that is a question for a lawyer in your country rather than an inference from an American headline."}]},{"t":"h2","id":"what-to-do-in-the-meantime","runs":[{"text":"What to do in the meantime"}]},{"t":"p","runs":[{"text":"If you build on generated material commercially: prefer providers offering indemnities and read what they cover; keep records of what was generated when and with what; avoid prompting for named living artists or protected characters; and treat any output that looks like something you recognise as a warning rather than a success."}]}],"check":{"question":[{"text":"A US ruling found that training a language model on books was transformative fair use, while separately finding that downloading a pirated library was not. What is the significance of that split?"}],"options":[{"text":[{"text":"It separates how copies were obtained from their use: fair use cannot cure piracy"}],"correct":true,"tempting":null},{"text":[{"text":"It establishes that training is lawful everywhere so long as the model is transformative"}],"correct":false,"tempting":[{"text":"Turns one district ruling into an international rule, when other judges in the same country declined to state any general rule."}]},{"text":[{"text":"Developers may train freely provided they delete the pirated copies afterwards"}],"correct":false,"tempting":[{"text":"Deleting after the fact does not undo the copying already done, which is the act the second holding addressed."}]},{"text":[{"text":"Fair use turns mainly on whether the works were published, and books are published"}],"correct":false,"tempting":[{"text":"Publication is part of the second fair use factor, but it was not what divided the two holdings."}]}],"order":[1,2,3,0],"why":[{"text":"Training was the use and the download was the acquisition. The court judged each separately."}]},"prev":{"slug":"what-copyright-actually-protects","title":"What copyright covers, and what it never did","minutes":8},"next":{"slug":"licences-you-actually-have","title":"Reading the licence before you build on it","minutes":8}}}